Dancheng MENG

Visiting scholar at NUS. Ph.D. at Tongji University. Neuro-architecture. Health Community and Healing Environment.

[ClaudeDevs @ClaudeDevs]

https://t.co/dxoI9Nna86

[Xudong Han @Xudong07452910]

现在很多人聊 Research Agent，默认期待是：

读论文、找 gap、想 idea、做实验、写 paper。

但耶鲁大学的这篇论文问了一个更深入的问题：

LLM 生成的研究想法，和人类研究者真正做出来的 paper idea，差距到底在哪里？

论文的核心概念叫 research taste，可以理解成「研究品味」：你通常会发现什么样的问题，用什么方式把它变成一个 contribution。

它的实验设计挺巧。

作者从真实论文里抽取 human idea，再反推 4 到 8 篇最可能启发这个 idea 的 prior works。然后把这些 prior works 的标题和摘要给 LLM，让模型在同样的文献上下文里生成一个新的 research idea。

这样比较的重点，就从「单个 idea 看起来新不新」推进到「一批 idea 的分布像不像人类研究者」。

论文一共用了 11,683 个 human ideas，覆盖 ML 会议论文和 Nature Communications 里的自然科学论文；模型包括 Claude、Gemini、GPT、Qwen、DeepSeek 等 9 个设置。

结果很有意思：人类 idea 的分布明显更宽。

在人类论文里，只有 12.1% 的 idea 属于 bridge opportunity，也就是把不同文献、方法或证据流连接起来。

但在 LLM 生成的 idea 里，这个比例变成了 47.1% 到 64.2%。

方法层面也类似。

人类论文里，synthesis / unification 只占 5.1%；LLM 这里是 22.5% 到 38.7%。

简单说，模型很喜欢把研究问题理解成：

这里有两个东西，可以把它们整合一下。

论文还发现，给模型更多上下文也没有明显解决这个问题。full-paper context 版本没有让分布更接近人类；开启 thinking mode 甚至可能让模型更强化这种 bridge-and-synthesis 倾向。

这篇最有意思的地方在于，它没有简单说 LLM idea 好或不好，而是指出了一个更隐蔽的问题：

LLM 可以生成很多看起来合理的研究想法，但这些想法的类型可能高度集中。

这对 AI for Science 和 Research Agent 很关键。

对于真正强的 Research Agent，难点在于能不能形成更宽的 problem-finding 能力：发现 failure、识别机制、提出测量工具、构造系统、修改局部假设，而不总是回到「整合两个已有东西」这条安全路径。

自动科研如果缺少这种多样性，最后可能会变成一种很流畅的同质化机器。

📎 arxiv: https://t.co/GrrulIuJrF

[Justin Skycak @justinskycak]

It's amazing, the level of student potential that gets unleashed, just by having them run this active learning loop on a knowledge graph.

[孤桜ETH @GYLQ520]

2017年，有位朋友向他请教“高效学英语的路径”，他当场梳理了一份实操指南。

时隔9年，这份指南演变成了一个斩获54.8k Star、5.6k Fork的开源项目，成了中文技术社区公认最硬核的英语学习资源库。

作者当年高考英语是江苏卷第一名，他的方法论直接把天花板拉高了一个维度——英语不是靠死记硬背堆出来的，而是像学母语一样在真实使用中自然内化。

词汇章节帮你彻底甩掉在Anki里刷屏式自我感动，听力部分手把手带着你一步步筑牢语感地基，口语篇把输出逻辑一层层拆开讲透，听、说、读、写整条链路全部打通。

而2026新版AI篇，才是真正让人眼前一亮的部分。

它不是教你拿AI当廉价翻译蒙混过关，而是把Gemini当作主引擎来驱动整个流程——Live用来练口语，Canvas拆解长难句，flashcards做闪卡记忆，ChatGPT分析语法结构，Perplexity查真实语境，DeepL Write精修输出表达。
🔗https://t.co/dQIXMjKaMJ

[大漂亮| C Labs @giantcutie666]

世行报告全文在此，我认为这是资本主义的邪路，本帖纯粹是学术性分享，不对世行建议正确与否作出任何评价😅
https://t.co/R6Mut2t604

[Vincent @Vincent_AINotes]

🔗 开源地址：https://t.co/tujFOxpel7

核心能力：
→ 论文语义向量化 → 3D空间映射
→ 搜索、缩放、旋转、聚类筛选
→ 悬停即可查看论文标题/领域/年份
→ 支持 Force Graph / 热力图 / 分布图等多种视图

适合谁：
📚 科研新手找方向
🎓 研究生写文献综述
🔬 研究员看跨学科连接
💡 任何想快速理解一个领域全貌的人

[🔻agitprop + absurdity🔻 @agtprpnabsrdty]

Hedge fund managers, venture capitalists, and tech executives are pulling their kids from public and private schools to enroll them in AI-tutoring startups that pay staff to stop calling themselves teachers and face no state requirement to report how well any of it works.

Alpha School, now in a dozen cities and expanding to nearly two dozen more this fall, charges $75,000 a year in San Francisco for two hours of AI tutoring followed by project workshops, using software that tracks how closely a child pays attention. Forge Prep in New Jersey charges up to $36,000 and promises graduates a $200,000 payout if they start a company instead of taking a job. Billionaire Bill Ackman has publicly backed Alpha. Neither school reports outcomes to any state, and Forge's own founder admits there is no data showing things are going well.

Alpha sells home-schooling software on top of tuition. Forge's startup fund assumes graduates will found companies rather than join a labor market its own AI curriculum is designed to help automate. Alpha's spokeswoman says its families are overwhelmingly finance, venture capital, and tech workers, meaning the industry building the AI disrupting entry-level jobs is now selling itself a private curriculum to land its own kids on the ownership side of that disruption.

Alpha's in-person staff voted to reject the word teacher, and its remote coaches monitoring the AI software are scattered across the globe. Stanford's Victor Lee says dropping the title diminishes the professionalism teaching requires. Stanford's Caroline Hoxby, who studies these programs rather than sells them, says there is negligible scientific evidence behind any of it.

My take: The same investor class funding AI tools built to gut entry-level jobs is now buying its own children a private exemption from those tools' consequences, for $75,000 a year.

Public schools absorb AI disruption unfunded and unreformed while this cohort builds itself an unaccountable alternative marketed as personalization. This is what class stratification looks like once it gets rebranded as coaching, guides, and life skills.

https://t.co/eAZnvCxex1

[ClaudeDevs @ClaudeDevs]

https://t.co/v694m8Eaj6

[Anthropic @AnthropicAI]

New Anthropic research: A global workspace in language models.

Of everything happening in your brain right now, only a tiny fraction is consciously accessible—thoughts you can describe, hold in mind, and reason with.

We found a strikingly similar divide inside Claude.

[Nicholas Fabiano, MD @NTFabiano]

Learning a second language earlier is associated with significantly greater brain connectivity & efficiency.

[Manthan Gupta @manthanguptaa]

https://t.co/CzmfXDgPRP

[Alex Finn @AlexFinn]

Yesterday I installed ClawdBot on this mac mini. An AI agent assistant that works for you 24/7

Since then it's accomplished all of this for me while I lived my life:

• Wrote 3 Youtube scripts
• Wrote my next newsletter
• Researched 26 other AI accounts and took notes on what's working
• Created a daily brief that has all the latest AI news
• Built it's own project management system
• Spun up it's own employees. I now have 2 levels of AI agent employees under me
• A complete 2nd brain system to replace Notion

Yeah. We literally have AGI

[OpenClaw🦞 @openclaw]

🦞Clawdbot 2026.1.24

📱 LINE channel with rich replies & quick actions
💬 Telegram DM topics = separate sessions now
✅ /approve exec prompts
✨ Control UI got a proper glow-up
📚 Ollama + Venice
🎙️ Edge TTS

The lobster provides😏
https://t.co/10mRJs11n6 https://t.co/BU0lcOhx3z

[Dancheng MENG @mengdancheng]

recommend🥰

[Emmanuel Tsekleves @PhDtoProf]

Simplify Your PhD Thesis & Dissertation Structure in 7 Clear Steps

Diving into a PhD thesis or dissertation can feel like navigating a labyrinth.

But what if we could break it down into manageable, clearly defined steps?

Here's how:

[Prof Lennart Nacke, PhD @acagamic]

You've spent months on your research, only to realize you used the wrong method. Ouch. Research methods confuse most people.

I see more boring, unappealing, and standard explanations by the day.

Because it's challenging to understand them all.

It requires many skills. Like:

→ Data analysis
→ Critical thinking
→ Experimental design
→ Statistical interpretation

All part of conducting great research.

Yet, many researchers fail at choosing the right method.

Here are 5 key branches of research methods:

→ Quantitative (numbers-based analysis)
→ Qualitative (exploratory, non-numerical)
→ Mixed Methods (combining both)
→ Experimental (testing cause-effect)
→ Observational (watching without interference)

A good researcher must be able to navigate these methods.

Qualitative methods are crucial for in-depth understanding.

Here's how I would structure a qualitative study:

1. Research question (what you want to explore)
2. Data collection (interviews, observations)
3. Thematic analysis (finding patterns)
4. Interpretation (making sense of findings)
5. Conclusion (addressing the research question)

I'd aim for rich, detailed descriptions in my methodology.

And for as much depth and context as my data allows.

Here are 7 qualitative methods I love:

1. Ethnography
2. Case Study
3. Phenomenology
4. Grounded Theory
5. Content Analysis
6. Action Research
7. Historical Research

Good luck with choosing a research method that doesn't fail.

It's a challenge. I think my methods could improve as well.

Please share your favourite research method in the comments if it helped you publish a paper.

It's helpful to many new researchers to get inspiration.

What do you think makes or breaks a good research methodology?

[Prof Lennart Nacke, PhD @acagamic]

7 steps of the scientific method structured by research type

[The Struggling Scientists @TheStrugglingS4]

The transformation is intense😅

[Easlo @heyeaslo]

My Productivity System for 2025:

[Easlo @heyeaslo]

My favorite iPhone apps right now: